Visualization for TW General Election

2020 Presidential and Legislative Results

Abstract

The notebook provides a nearly-neutral visual description on the 2020 Taiwanese presidential and legislative election results. The polling took place at 2020-01-11 and data has been collected by Central Election Commision (CEC) at the polling place level.

---

The notebook is more of a visualization workshop using polling data as working examples. It is NOT meant to provide any in-depth political insights.

---

Prerequisites

Platform

The source data contains traditional Chinese characters encoded in UTF-8. To avoid troublesome encoding issues (especially in a Windows platform with a default CP1252 code page), the codes in this notebook are only tested on Ubuntu and macOS. Codes are in general still working under Windows. Just that some code chunks may not be able to print out unicode characters nicely.

Data

Election Data

We use minimally pre-processed data for our analytics. Please refer to the repo TW_Presidential_Election_2020 for details about the pre-processing logic. To directly download the archives please visit here. We put the uncompressed files under data/processed at our project root directory.

Map Data

For Taiwan map information, the shp file (shapefile) can be downloaded at GADM.org. We put the unzipped files under data/map at our project root directory.

Dependencies

Additional dependencies are required for Linux machine in order to do map plotting. Here is the instruction for Ubuntu.

Install gdal:

sudo add-apt-repository ppa:ubuntugis/ubuntugis-unstable
sudo apt update
sudo apt install libgdal-dev

Install ubunits:

sudo apt install libudunits2-dev

After installation of the above packages, we can then install and import the following R packages for this project:

library(data.table)

# For viualization.
library(ggplot2)
library(ggrepel)
library(plotly)

# Fix macOS ggplot device to locate a Chinese font.
# To make `knitr` aware of this we also need to speficy `fig.showtext=TRUE` for chunks with ggplot output.
library(showtext)
font_add_google("Noto Serif", "noto")
showtext_auto()

# For spatial data visualization.
library(sf)

# For linear modeling with robust std err.
library(lmtest)
library(sandwich)

Background

• 2020 15th Taiwanese Presidential Election
• 2020 10th Taiwanese Legislative Election

Visual Analytics on Election Results

第十五屆總統選舉結果 The 15th Presidential Election Results

有效票 Valid Vote Counts

By County 依縣市

# Read shapefile for spatial information about counties in Taiwan.
tw_county_sf <- sf::st_read("data/map/gadm36_TWN_2.shp", quiet=TRUE)

# Align county naming.
tw_county_sf$NL_NAME_2 <- gsub("台中", "台中市", tw_county_sf$NL_NAME_2)
tw_county_sf$NL_NAME_2 <- gsub("台南", "台南市", tw_county_sf$NL_NAME_2)
tw_county_sf$NL_NAME_2 <- gsub("馬祖列島", "連江縣", tw_county_sf$NL_NAME_2)

# process county-level presidential election results.
p_county <- fread("data/processed/presidential/presidential_counties.csv", encoding="UTF-8")
p_county <- p_county[, Region:=trimws(Region, whitespace="[\\h\\v]")]  # Trim full-width white space.
p_county <- p_county[, Region:=gsub("臺", "台", Region)]  # Align character with shapefile.
setnames(p_county, "投票率H H=C÷G", "Turnout")

# Final counts.
cols_candidate <- c("(1) 宋楚瑜 余湘", "(2) 韓國瑜 張善政", "(3) 蔡英文 賴清德")
p_final <- as.data.table(colSums(p_county[, cols_candidate, with=FALSE]), keep.rownames=TRUE)
setnames(p_final, c("Candidate", "Votes"))
p_final <- p_final[, Share:=Votes / sum(Votes)]

# Merge with shp to get English county name.
p_county <- merge(p_county, tw_county_sf[, c("NL_NAME_2", "NAME_2")],
                  by.y="NL_NAME_2", by.x="Region")
p_county$geometry <-NULL
setnames(p_county, "NAME_2", "RegionEn")
p_county <- p_county[, County:=paste(Region, RegionEn)]

# Melt from wide to long for plotting purpose.
cols_region <- c("County", "Region", "RegionEn")
p_county_l <- melt(p_county[, c(cols_region, cols_candidate), with=FALSE],
                   id.vars=cols_region, variable.name="Candidate", value.name="Votes")
p_county_l <- p_county_l[, Candidate:=factor(Candidate)]  # Level by candidate number.

# Compute difference between DPP and KMT by county.
setorder(p_county_l, County, Candidate)
p_county_l <- p_county_l[, Votes_Diff:=Votes[3] - Votes[2], by=.(County)]

# Compute total votes by county.
p_county_l <- p_county_l[, Total_Votes:=sum(Votes), by=.(County)]

# Compute share by county.
p_county_l <- p_county_l[, Votes_Share:=Votes / Total_Votes]
p_county_l <- p_county_l[, Share_Diff:=Votes_Share[3] - Votes_Share[2], by=.(County)]
p_county_diff <- unique(p_county_l, by="County")[
  , c(cols_region, "Total_Votes", "Share_Diff"), with=FALSE]
p_county_diff <- p_county_diff[, dpp_win:=Share_Diff > 0]

# Join share diff to spatial dataframe.
tw_county_sf <- merge(tw_county_sf, p_county_diff, by.x="NL_NAME_2", by.y="Region")

# Compute difference between DPP and KMT on the wide data as well.
p_county <- p_county[, Vote_Diff:=get("(3) 蔡英文 賴清德") - get("(2) 韓國瑜 張善政")]
p_county <- p_county[, Share_Diff:=Vote_Diff / get("有效票數A A=1+2+...+N")]

# Color configuration for main parties.
main_party_colors <- c("darkorange", "dodgerblue", "green3")

ggplot(p_final, aes(x=Candidate, y=Votes, fill=Candidate)) +
  geom_bar(stat="identity") +
  labs(x="候選人", y="得票數",
       title="2020台灣總統大選結果") +
  geom_text(aes(x=Candidate, y=Votes / 2,
                label=sprintf("%s\n(%s)", scales::comma(Votes), scales::percent(Share))),
            size=5) +
  scale_fill_manual(values=main_party_colors) +
  scale_y_continuous(labels=scales::comma) +
  theme(legend.position="none") +
  coord_flip()

2020 Presidential Election Final Outcome

The pre-processed county result for presidential election is a wide format by county:

head(p_county[, -c(1:3)])

And we’ve reshaped it into a long format for ease of plotting:

# Show result for Taipei.
p_county_l[Region == "台北市"]

Plot vote counts by county:

p <- ggplot(p_county_l, aes(x=reorder(County, Total_Votes), y=Votes, fill=Candidate)) +
  geom_bar(stat="identity", position="dodge", width=.8) +
  scale_fill_manual(name="候選人", values=main_party_colors) +
  scale_y_continuous(labels=scales::comma) +
  labs(x="縣市（依總票數排序）", y="得票數",
       title="2020台灣總統大選候選人各縣市：得票數") +
  coord_flip() +
  theme(legend.position="bottom")

ggplotly(p, width=800, height=600)

And also vote shares (vote counts / total valid votes):

p <- ggplot(p_county_l, aes(x=reorder(County, Total_Votes), y=Votes_Share, fill=Candidate)) +
  geom_bar(stat="identity", position="stack", width=.8) +
  geom_text(aes(label=scales::percent(Votes_Share)), position=position_stack(vjust=.5), size=2) +
  scale_fill_manual(name="候選人", values=main_party_colors) +
  labs(x="縣市（依總票數排序）", y="得票率",
       title="2020台灣總統大選候選人各縣市：得票率") +
  coord_flip() +
  theme(legend.position="bottom")

ggplotly(p, width=800, height=600)

Since the main parties are DDP and KMT, we define vote share difference as vote share of DDP minus that of KMT. The following barplot gives us a clear view of which county DPP won the most.

p <- ggplot(p_county_diff, aes(x=reorder(County, Share_Diff), fill=dpp_win, y=Share_Diff)) +
  geom_bar(stat="identity", width=.8, color="black", alpha=.5) +
  scale_fill_manual(values=c("dodgerblue", "green3")) +
  theme(legend.position="none") +
  geom_text(aes(label=scales::percent(Share_Diff)), position=position_stack(vjust=.5), size=2.5) +
  labs(x="縣市（依得票率差排序）", y="綠藍得票率差",
       title="2020台灣總統大選各縣市：綠藍得票率差") +
  coord_flip()

ggplotly(p, width=800, height=600)

We can also visualize the above ranking with a choropleth map:

p <- ggplot(tw_county_sf, aes(text=County)) +
  geom_sf(aes(fill=Share_Diff)) +
  coord_sf(xlim=c(118, 122.5), ylim=c(21.5, 26.5)) +
  scale_fill_gradient(name="綠藍得票率差", low="blue3", high="green3") +
  labs(title="2020台灣總統大選各縣市：綠藍得票率差")

ggplotly(p, width=800, height=600)

Apparently DPP won most counties besides maintain and outlying island areas.

Go Back to Top

By County Sub-Region 依鄉鎮市區

# Read county subregion-level presidential election results.
p_region <- fread("data/processed/presidential/presidential_regions.csv", encoding="UTF-8")
p_region <- p_region[, Region:=trimws(Region, whitespace="[\\h\\v]")]  # Trim full-width white space.
setnames(p_region, "投票率H H=C÷G", "Turnout")

# Compute difference between DPP and KMT by region.
p_region <- p_region[, Vote_Diff:=get("(3) 蔡英文 賴清德") - get("(2) 韓國瑜 張善政")]
p_region <- p_region[, Share_Diff:=Vote_Diff / get("有效票數A A=1+2+...+N")]

# Wide to long.
p_region_l <- melt(p_region[, c("By", "Region", cols_candidate), with=FALSE],
                   id.vars=c("By", "Region"), variable.name="Candidate", value.name="Votes")
p_region_l <- p_region_l[, Candidate:=factor(Candidate)]  # Level by candidate number.
p_region_l <- p_region_l[p_region, Share_Diff:=i.Share_Diff, on=.(Region)]

setorder(p_region, -Share_Diff)

# Group counties into areas.
north_counties <- c("臺北市", "新北市", "基隆市", "新竹市", "桃園市", "新竹縣", "宜蘭縣")
central_counties <- c("臺中市", "苗栗縣", "彰化縣", "南投縣", "雲林縣")
south_counties <- c("高雄市", "臺南市", "嘉義市", "嘉義縣", "屏東縣", "澎湖縣")
east_counties <- c("花蓮縣", "臺東縣")
island_counties <- c("金門縣", "連江縣")

# Helper function for share difference bar plot.
plot_region_share_diff <- function(counties, title) {
  DT <- p_region[By %in% counties]
  DT[, By:=factor(By, levels=counties)]  # Fix order for facet grids.
  DT[, Region2:=paste(By, Region)]  # Fixed order within facet (same-name region for different counties).
  DT[, dpp_win:=Share_Diff > 0]
  p <- ggplot(DT, aes(x=reorder(Region2, Share_Diff), y=Share_Diff, fill=dpp_win)) +
    geom_bar(stat="identity", width=.8, color="black", alpha=.5) +
    scale_fill_manual(values=c("dodgerblue", "green3")) +
    theme(legend.position="none") +
    geom_text(aes(label=scales::percent(Share_Diff)), position=position_stack(vjust=.5), size=2.5) +
    labs(x="", y="", title=title) +
    coord_flip() +
    facet_grid(By ~ ., scales="free_y", space="free_y")
  p
}

Vote share difference (DPP v.s. KMT) ranked by region across the nation:

p_region_rank <- p_region[, .(By, Region, Share_Diff, Vote_Diff)]
p_region_rank <- p_region_rank[, Share_Diff:=scales::percent(Share_Diff, .001)]
p_region_rank <- p_region_rank[, i:=.I]
setcolorder(p_region_rank, "i")
setnames(p_region_rank, c("排序", "縣市", "鄉鎮市區", "綠藍得票率差", "綠藍得票數差"))
p_region_rank[]

We bar-plot the above ranking by region by county:

North Taiwan 北台灣

plot_region_share_diff(
  north_counties,
  title="2020台灣總統大選北部地區（鄉鎮市區）：綠藍得票率差")

Central Taiwan 中台灣

plot_region_share_diff(
  central_counties,
  title="2020台灣總統大選中部地區（鄉鎮市區）：綠藍得票率差")

South Taiwan 南臺灣

plot_region_share_diff(
  south_counties,
  title="2020台灣總統大選南部地區（鄉鎮市區）：綠藍得票率差")

East Taiwan 東台灣

plot_region_share_diff(
  east_counties,
  title="2020台灣總統大選東部地區（鄉鎮市區）：綠藍得票率差")

Outlying Islands 離島

plot_region_share_diff(
  island_counties,
  title="2020台灣總統大選離島地區（鄉鎮市區）：綠藍得票率差")

無效票 Invalid Vote Counts

Invalid votes are votes cast but declared invalid. There is no way to objectively record the intention of an invalid vote, here we simply statistically describe the share of invalid votes from total votes cast.

By County 依縣市

# Process invalid votes by county.
p_county <- p_county[, Invalid_Votes_Share:=get("無效票數B") / get("投票數C C=A+B")]
p_county <- p_county[, Invalid_Votes_ShareP:=Invalid_Votes_Share * 100]

# Join invalid vote shares to spatial dataframe.
tw_county_sf <- merge(tw_county_sf, p_county[, .(Region, Invalid_Votes_ShareP)], by.x="NL_NAME_2", by.y="Region")

p <- ggplot(p_county, aes(x=reorder(County, Invalid_Votes_Share), y=Invalid_Votes_ShareP)) +
  geom_bar(stat="identity", width=.8, fill="white", color="black") +
  geom_text(aes(label=scales::percent(Invalid_Votes_Share)), position=position_stack(vjust=.5), size=2.5) +
  labs(x="縣市（依廢票率排序）", y="廢票率％",
       title="2020台灣總統大選各縣市：廢票率") +
  coord_flip()

ggplotly(p, width=800, height=600)

To see the geolocational differences:

p <- ggplot(tw_county_sf, aes(text=County)) +
  geom_sf(aes(fill=Invalid_Votes_ShareP)) +
  coord_sf(xlim=c(118, 122.5), ylim=c(21.5, 26.5)) +
  scale_fill_gradient(name="廢票率％", low="white", high="black") +
  labs(title="2020台灣總統大選各縣市：廢票率％")

ggplotly(p, width=800, height=600)

Check if there is any correlation between invalid votes and the DPP-KMT differences:

ggplot(p_county, aes(x=Invalid_Votes_Share, y=Share_Diff, label=County)) +
  geom_point(shape=4, size=4, stroke=2) +
  labs(x="廢票率", y="綠藍得票率差",
       title="2020台灣總統大選各縣市：廢票率Ｘ綠藍得票率差") +
  geom_text_repel(color="grey40")

2020 Presidential Election Invalid Vote Shares vs DPP-KMT Share Differences by County

Go Back to Top

By County Sub-Region 依鄉鎮市區

# Process invalid votes by region.
p_region <- p_region[, Invalid_Votes_Share:=get("無效票數B") / get("投票數C C=A+B")]
p_region <- p_region[, Invalid_Votes_ShareP:=Invalid_Votes_Share * 100]

p <- ggplot(p_region, aes(x=Invalid_Votes_Share, y=Share_Diff, label=paste(By, Region))) +
  geom_point(shape=4, size=4) +
  labs(x="廢票率", y="綠藍得票率差",
       title="2020台灣總統大選各鄉鎮市區：廢票率Ｘ綠藍得票率差") +
       geom_smooth(method="lm", color="black")
ggplotly(p, width=800, height=600)

A simple linear model on invalid vote shares against DPP-KMT share differences suggests the correlation is at least statistically significant. But such correlation may be spurious so the interpretation shall be cautious.

lm_invalid <- lm(Share_Diff ~ Invalid_Votes_ShareP, data=p_region)
coeftest(lm_invalid, vcov=vcovHC(lm_invalid, type="HC0"))

t test of coefficients:

                      Estimate Std. Error t value Pr(>|t|)   
(Intercept)          -0.196475   0.103482 -1.8986 0.058399 . 
Invalid_Votes_ShareP  0.276780   0.086798  3.1888 0.001552 **
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Go Back to Top

投票率 Voter Turnout

Voter turnout is defined as total votes cast (valid and invalid) divided by legitimate voters. The information is readily available in the source data from the column 投票率H H=C÷G. In this section we explore the possible relation between turnout and vote counts.

By County 依縣市

# Get voting rate from wide data to long.
p_county_l <- p_county_l[p_county, Turnout:=i.Turnout, on=.(County)]

p <- ggplot(p_county, aes(x=reorder(County, Turnout), y=Turnout)) +
  geom_bar(stat="identity", width=.8, fill="white", color="black") +
  geom_text(aes(label=paste(Turnout, "%")), position=position_stack(vjust=.5), size=2.5) +
  labs(x="縣市（依投票率排序）", y="投票率",
       title="2020台灣總統大選各縣市：投票率") +
  coord_flip()

ggplotly(p, width=800, height=600, tooltip=NULL)

It is widely suggested that an increasing voter turnout due to younger generation this time can hugely impact the results. Here we plot the turnout against DPP-KMT share differences by county:

ggplot(p_county, aes(x=Turnout, y=Share_Diff, label=County)) +
  scale_x_continuous(limits=c(35, 80)) +
  geom_rect(data=p_county[1],
            aes(xmin=-Inf, xmax=Inf, ymin=0, ymax=Inf), fill="lightgreen", alpha=.8) +
  geom_rect(data=p_county[1],
            aes(xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=0), fill="lightblue", alpha=.8) +
  geom_hline(aes(yintercept=0), color="red") +
  geom_point() +
  geom_smooth(method="lm") +
  geom_text(check_overlap=TRUE, color="grey40") +
  labs(x="投票率％", y="綠藍得票率差",
       title="2020台灣總統大選各縣市：投票率Ｘ藍綠得票率差")

2020 Presidential Election Voter Turnout by County

We zoom-in in the following plot for those with turnout > 70%:

ggplot(p_county[Turnout > 70], aes(x=Turnout, y=Share_Diff, label=County)) +
  scale_x_continuous(limits=c(70, 80)) +
  geom_rect(data=p_county[Turnout > 70][1],
            aes(xmin=-Inf, xmax=Inf, ymin=0, ymax=Inf), fill="lightgreen", alpha=.8) +
  geom_rect(data=p_county[Turnout > 70][1],
            aes(xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=0), fill="lightblue", alpha=.8) +
  geom_hline(aes(yintercept=0), color="red") +
  geom_point() +
  geom_text_repel(color="grey40") +
  labs(x="投票率％", y="綠藍得票率差",
       title="2020台灣總統大選各縣市：投票率(>70%)Ｘ綠藍得票率差")

2020 Presidential Election Voter Turnout (>70%) by County

We can also visualize the turnout relation on each party candidate instead:

ggplot(p_county_l, aes(x=Turnout, y=Votes_Share)) +
  geom_rect(data=unique(p_county_l, by="Candidate"), aes(fill=Candidate),
            xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=Inf, alpha=.5) +
  scale_fill_manual(values=main_party_colors) +
  theme(legend.position="none") +
  geom_point() +
  geom_smooth(method="lm", color="black") +
  facet_grid(~ Candidate) +
  labs(x="投票率％", y="得票率",
       title="2020台灣總統大選各縣市：投票率Ｘ得票率")

2020 Presidential Election Voter Turnout vs Candidate Vote Shares by County

Go Back to Top

By County Sub-Region 依鄉鎮市區

# Compute region-level voting rate and vote shares in the long data.
p_region_l <- p_region_l[p_region, Turnout:=i.Turnout, on=.(Region)]
p_region_l <- p_region_l[, Votes_Share:=Votes / sum(Votes), by=.(Region)]

From the previous county-level visualization we realized that turnout may be related to the difference between DDP and KMT candidates. Here we use a simple linear model to test the correlation based on region-level data:

lm_turnout <- lm(Share_Diff ~ Turnout, data=p_region_l)
coeftest(lm_turnout, vcov=vcovHC(lm_turnout, type="HC0"))

t test of coefficients:

              Estimate Std. Error t value  Pr(>|t|)    
(Intercept) -1.2211187  0.0882236 -13.841 < 2.2e-16 ***
Turnout      0.0185861  0.0012014  15.470 < 2.2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Apparently the correlation is significant. Let’s also generate the region-level scatterplot:

ggplot(p_region, aes(x=Turnout, y=Share_Diff)) +
  geom_rect(data=p_region[1], aes(xmin=-Inf, xmax=Inf, ymin=0, ymax=Inf),
            fill="lightgreen", alpha=.8) +
  geom_rect(data=p_region[1], aes(xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=0),
            fill="lightblue", alpha=.8) +
  geom_hline(aes(yintercept=0), color="red") +
  geom_point() +
  geom_smooth(method="lm") +
  labs(x="投票率％", y="綠藍得票率差",
       title="2020台灣總統大選各鄉鎮市區：投票率Ｘ綠藍得票率差")

2020 Presidential Election Voter Turnout by Region

Again with additional zoom-in:

ggplot(p_region[Turnout > 70], aes(x=Turnout, y=Share_Diff, label=Region)) +
  geom_rect(data=p_region[1], aes(xmin=-Inf, xmax=Inf, ymin=0, ymax=Inf),
            fill="lightgreen", alpha=.8) +
  geom_rect(data=p_region[1], aes(xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=0),
            fill="lightblue", alpha=.8) +
  geom_hline(aes(yintercept=0), color="red") +
  geom_point() +
  geom_smooth(method="lm", se=FALSE) +
  geom_text(check_overlap=TRUE, nudge_y=-.035, color="darkgrey") +
  labs(x="投票率％", y="綠藍得票率差",
       title="2020台灣總統大選各鄉鎮市區：投票率(>70%)Ｘ綠藍得票率差")

2020 Presidential Election Voter Turnout (>70%) by Region

We can also test the correlation between turnout and vote share of each party’s candidates:

for ( candidate in unique(p_region_l$Candidate) ) {
  lm_turnout_candidiate <- lm(Votes_Share ~ Turnout, data=p_region_l[Candidate == candidate])
  message(rep("-", 80))
  message("Test correlation between vote shares and voter turnout.")
  message(sprintf("For %s:", candidate))
  print(coeftest(lm_turnout_candidiate, vcov=vcovHC(lm_turnout_candidiate, type="HC0")))
}

--------------------------------------------------------------------------------

Test correlation between vote shares and voter turnout.

For (1) 宋楚瑜 余湘:

t test of coefficients:

              Estimate Std. Error t value  Pr(>|t|)    
(Intercept) 3.5845e-02 9.1138e-03  3.9331 0.0001003 ***
Turnout     1.8894e-05 1.2509e-04  0.1510 0.8800239    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

--------------------------------------------------------------------------------

Test correlation between vote shares and voter turnout.

For (2) 韓國瑜 張善政:

t test of coefficients:

             Estimate Std. Error t value  Pr(>|t|)    
(Intercept)  1.135919   0.081971  13.857 < 2.2e-16 ***
Turnout     -0.010086   0.001121  -8.997 < 2.2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

--------------------------------------------------------------------------------

Test correlation between vote shares and voter turnout.

For (3) 蔡英文 賴清德:

t test of coefficients:

              Estimate Std. Error t value  Pr(>|t|)    
(Intercept) -0.0695098  0.0737391 -0.9426    0.3465    
Turnout      0.0082257  0.0010196  8.0673 1.039e-14 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Interestingly, such correlation is not found for PFP candidates.

ggplot(p_region_l, aes(x=Turnout, y=Votes_Share)) +
  geom_rect(data=unique(p_region_l, by="Candidate"), aes(fill=Candidate),
            xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=Inf, alpha=.5) +
  scale_fill_manual(values=main_party_colors) +
  theme(legend.position="none") +
  geom_point() +
  geom_smooth(method="lm", color="black") +
  facet_grid(~ Candidate) +
  labs(x="投票率％", y="得票率",
       title="2020台灣總統大選各縣市：投票率 X 得票率")

2020 Presidential Election Voter Turnout vs Candidate Vote Shares by Region

Go Back to Top

By Polling Place 依投開票所

TBC.

Go Back to Top

第十屆立法委員選舉結果 The 10th Legislative Election Results

不分區立委 Party List

# Process party-list results.
# Note that the table is down to polling place level.
partylist <- fread("data/processed/legislative/legislative_partylist.csv", encoding="UTF-8")
partylist <- partylist[, Region:=trimws(Region, whitespace="[\\h\\v]")]  # Trim full-width white space.
cols_area <- c("By", "Region", "Village", "Place")
cols_party <- names(partylist)[7:25]  # There are 19 parties in total.
partylist_l <- melt(partylist[, c(cols_area, cols_party), with=FALSE],
                   id.vars=cols_area, variable.name="Party", value.name="Votes")
partylist_l <- partylist_l[, Party:=factor(Party)]
partylist_final <- partylist_l[, .(Votes=sum(Votes)), by=.(Party)]
partylist_final <- partylist_final[, Share:=Votes / sum(Votes)]

For party list to secure at least one seat in Legislative Yuan, it must obtain minimally 5% of party-list votes. If a party cannot pass the 5% threshold but pass the 3% threshold, it will be sponsored by the government in the following 4 years (based on number of votes obtained). This is particularly important for small parties to cover their operating costs.

thres_003 <- sum(partylist_final$Votes) * .03
thres_005 <- sum(partylist_final$Votes) * .05

ggplot(partylist_final, aes(x=reorder(Party, Votes), y=Votes)) +
  geom_bar(stat="identity") +
  labs(x="政黨", y="得票數",
       title="2020台灣不分區立委選舉結果") +
  geom_text(aes(x=Party, y=Votes,
                label=sprintf("%s (%s)", scales::comma(Votes), scales::percent(Share, .01))),
            size=3, hjust=-.05) +
  scale_y_continuous(labels=scales::comma, limits=c(0, max(partylist_final$Votes) * 1.25)) +
  geom_hline(aes(yintercept=thres_003, color="補助 Sponsorship"), linetype=3) +
  geom_hline(aes(yintercept=thres_005, color="席次 Seats"), linetype=3) +
  scale_colour_manual(name="門檻 Threshold", values=c("red", "green")) +
  theme(legend.position="bottom") +
  coord_flip()

2020 Legislative Election Party List Final Outcome

By County 依縣市

We group parties with less than 2% votes into “others” to make the rest of the plots easier to read.

partylist_county <- partylist_l[, .(Votes=sum(Votes)), by=.(By, Party)]
small_parties <- partylist_final[Share < .02, Party]

partylist_county <- partylist_county[Party %in% small_parties, Party:="其他"]
partylist_county <- partylist_county[, Party:=factor(Party)]  # Renew level.
partylist_county <- partylist_county[, .(Votes=sum(Votes)), by=.(By, Party)]  # Re-group.
partylist_county <- partylist_county[, Share:=Votes / sum(Votes), by=.(By)]

# Compute DPP-KMT share differences.
partylist_county_gb <- partylist_county[Party %in% c("(14)  民主進步黨", "(9)  中國國民黨")]
setorder(partylist_county_gb, By, Party)
partylist_county_gb_diff <- partylist_county_gb[, .(Share_Diff=Share[2] - Share[1]), by=.(By)]
partylist_county_gb_diff <- partylist_county_gb_diff[, dpp_win:=Share_Diff > 0]

party_colors <- c("darkorange", "darkgreen", "yellow", "dodgerblue", "green", "green3", "lightblue", "grey")
ggplot(partylist_county, aes(x=reorder(By, Votes), y=Share, fill=Party)) +
  geom_bar(stat="identity", position="stack", width=.8) +
  geom_text(aes(label=scales::percent(Share, .01)), position=position_stack(vjust=.5), size=2, angle=20) +
  scale_fill_manual(name="政黨", values=party_colors) +
  labs(x="縣市（依總票數排序）", y="得票率",
       title="2020台灣不分區立委各縣市政黨得票率") +
  coord_flip() +
  theme(legend.position="bottom")

2020 Legislative Election Party List Vote Shares by County

p <- ggplot(partylist_county_gb_diff, aes(x=reorder(By, Share_Diff), fill=dpp_win, y=Share_Diff)) +
  geom_bar(stat="identity", width=.8, color="black", alpha=.5) +
  scale_fill_manual(values=c("dodgerblue", "green3")) +
  theme(legend.position="none") +
  geom_text(aes(label=scales::percent(Share_Diff)), position=position_stack(vjust=.5), size=2.5) +
  labs(x="縣市（依得票率差排序）", y="綠藍得票率差",
       title="2020台灣不分區立委各縣市：綠藍得票率差") +
  coord_flip()

ggplotly(p, width=800, height=600)

Go Back to Top

By County Sub-Region 依鄉鎮市區

# Process region-level party list results.
partylist_region <- partylist_l[, .(Votes=sum(Votes)), by=.(By, Region, Party)]
partylist_region <- partylist_region[, Party:=factor(Party)]  # Renew level.
partylist_region <- partylist_region[, Share:=Votes / sum(Votes), by=.(By, Region)]
partylist_region <- partylist_region[!Party %in% small_parties]  # Ignore small parties.

# Get presidential invalid vote share and turnout.
partylist_region <- partylist_region[p_region, `:=`(
  P_Invalid_Votes_Share=i.Invalid_Votes_Share,
  P_Turnout=i.Turnout
  ), on=.(By, Region)]

# Get presidential vote results.
p_region_l <- p_region_l[, P_Share:=Votes / sum(Votes), by=.(By, Region)]
partylist_region <- merge(partylist_region, p_region_l[, .(By, Region, Candidate, P_Share, Share_Diff)], by=c("By", "Region"), allow.cartesian=TRUE)

# Helper function for scatter plot.
scatterplot <- function() {
  ggplot(partylist_region, aes(x=Share, y=P_Share, color=Candidate)) +
    geom_rect(data=unique(partylist_region, by=c("Party", "Candidate")), aes(fill=Party),
              xmin=-Inf, xmax=Inf, ymin=-Inf, ymax=Inf, alpha=.25) +
    scale_fill_manual(values=party_colors) +
    geom_point(size=.5) +
    geom_smooth(method="lm", aes(color=Candidate)) +
    scale_color_manual(values=main_party_colors) +
    labs(x="總統候選人得票率", y="不分區政黨得票率",
         title="2020總統候選人得票率Ｘ不分區政黨得票率") +
    facet_wrap(Party ~ Candidate, ncol=3, scales="free")
}

We can plot the correlation between presidential candidate vote shares and party vote shares at region level. But we must be very careful in interpreting the relation since we don’t know the direction of such relation, if any. Strictly speaking, we don’t even know whether they are causally related at all.

scatterplot()

2020 Legislative Election Party List Shares v.s. Presidential Votes

Go Back to Top

By Polling Place 依投開票所

TBC.

Go Back to Top

分區立委 Constituency

Overall 概括

# Process constityency results.
# Note that each district can have a varying number of candidates.
infiles_constituency <- grep(
  "constituency",dir("data/processed/legislative/", full.names=TRUE), value=TRUE)

c_all <- lapply(infiles_constituency, fread, encoding="UTF-8")
names(c_all) <- gsub("legislative_constituency_", "",
                     tools::file_path_sans_ext(basename(infiles_constituency)))

# Convert all table from wide to long.
wide_to_long <- function(DT) {
  candidate_start_ind <- 7
  candidate_end_ind <- ncol(DT) - 8
  cols_candidate <- names(DT)[candidate_start_ind:candidate_end_ind]
  cols_region <- c("By", "Region", "Village", "Place")
  out <- melt(DT[, c(cols_region, cols_candidate), with=FALSE],
              id.vars=cols_region, variable.name="Candidate", value.name="Votes")
  out
}

c_l <- rbindlist(lapply(c_all, wide_to_long))
c_l <- c_l[, Region:=trimws(Region, whitespace="[\\h\\v]")]  # Trim full-width white

# Extract candidate party for all districts.
c_l <- c_l[, Party:=gsub(".*[ ]", "", Candidate)]

# Sum by party.
c_party <- c_l[, .(Votes=sum(Votes)), by=.(Party)]
c_party <- c_party[, Share:=Votes / sum(Votes)]

Check number of polling places for each district:

c_pplace_cnt <- c_l[, .N, by=.(By)]
setorder(c_pplace_cnt, By)
setnames(c_pplace_cnt, c("District", "Num of Polling Place"))
c_pplace_cnt[]

ggplot(c_party, aes(x=reorder(Party, Votes), y=Votes)) +
  geom_bar(stat="identity") +
  labs(x="政黨", y="得票數",
       title="2020台灣分區立委選舉結果:依候選人政黨") +
  geom_text(aes(x=Party, y=Votes,
                label=sprintf("%s (%s)", scales::comma(Votes), scales::percent(Share, .01))),
            size=3, hjust=-.05) +
  scale_y_continuous(labels=scales::comma, limits=c(0, max(c_party$Votes) * 1.2)) +
  coord_flip()

2020 Legislative Election Constituency Outcome by Party

Go Back to Top

By County 依縣市

# Keep only the top few parties.
c_top_party <- c_party[Share > .01, Party]
c_party_colors <- c("dodgerblue", "green3", "black", "yellow", "lightblue", "darkgreen", "darkgrey")

# Sum by party by county.
c_l <- c_l[, County:=gsub("(第[0-9]+)?選舉區", "", By)]
c_party_county <- copy(c_l)
c_party_county <- c_party_county[!Party %in% c_top_party, Party:="其他"]
c_party_county <- c_party_county[, .(Votes=sum(Votes)), by=.(County, Party)]
c_party_county <- c_party_county[, Share:=Votes / sum(Votes), by=.(County)]
c_party_county[, Party:=factor(Party, level=c(c_top_party, "其他"))]

# Order county by total votes.
county_v_cnt <- c_l[, .(Votes=sum(Votes)), by=.(County)]
setorder(county_v_cnt, -Votes)
c_party_county <- c_party_county[, County:=factor(County, county_v_cnt$County)]

# Compute DPP-KMT share differences.
c_party_county_gb <- c_party_county[Party %in% c("民主進步黨", "中國國民黨")]
setorder(c_party_county_gb, County, Party)
# Note that DPP has no candidate for 1 county. We count the share as 0.
c_party_county_gb_diff <- c_party_county_gb[
  , .(Share_Diff=ifelse(is.na(Share[2]), 0, Share[2]) - Share[1]), by=.(County)]
c_party_county_gb_diff <- c_party_county_gb_diff[, dpp_win:=Share_Diff > 0]

We plot all parties with more than 1% vote share by county:

ggplot(c_party_county, aes(x=reorder(Party, Votes), y=Share, fill=Party)) +
  geom_bar(stat="identity", position="dodge", width=.75) +
  geom_text(aes(label=scales::percent(Share, .01)),
            position=position_dodge(.9), size=3, hjust=-.1) +
  scale_y_continuous(limits=c(0, 1), breaks=seq(0, 1, .2)) +
  scale_fill_manual(name="政黨", values=c_party_colors) +
  labs(x="政黨", y="得票率",
       title="2020台灣分區立委各縣市政黨得票率") +
  coord_flip() +
  theme(legend.position="none") +
  facet_wrap(County ~ ., ncol=5)

2020 Legislative Election Constituency Shares by Party by County

Also examine the DPP-KMT difference for constituency votes by county:

p <- ggplot(c_party_county_gb_diff, aes(x=reorder(County, Share_Diff), fill=dpp_win, y=Share_Diff)) +
  geom_bar(stat="identity", width=.8, color="black", alpha=.5) +
  scale_fill_manual(values=c("dodgerblue", "green3")) +
  theme(legend.position="none") +
  geom_text(aes(label=scales::percent(Share_Diff)), position=position_stack(vjust=.5), size=2.5) +
  labs(x="縣市（依得票率差排序）", y="綠藍得票率差",
       title="2020台灣分區立委各縣市：綠藍得票率差") +
  coord_flip()

ggplotly(p, width=800, height=600)

Go Back to Top

Taichung 3rd District 台中市第三選區

# Consolidate both party-list and constituency results for Taichung district 3.
tc3 <- c_l[By == "臺中市第3選舉區"]
tc3 <- tc3[, C_Share:=Votes / sum(Votes)]

# Compute party-list vote shares by polling place.
tc3_pl <- partylist_l[By == tc3$County[1] & Region %in% tc3$Region]
tc3_pl <- tc3_pl[, Share:=Votes / sum(Votes), by=.(Place)]
tc3_candidates <- c("(3) 洪慈庸 無", "(4) 張睿倉 台灣民眾黨", "(6) 楊瓊瓔 中國國民黨")
tc3_final <- merge(tc3_pl, tc3[Candidate %in% tc3_candidates,
                               .(Place, Candidate, C_Share)],
                   by="Place", allow.cartesian=TRUE)

tc3r <- tc3[, .(Votes=sum(Votes)), by=.(Candidate)]
tc3r <- tc3r[, Share:=Votes / sum(Votes)]
ggplot(tc3r, aes(x=reorder(Candidate, Votes), y=Votes)) +
  geom_bar(stat="identity") +
  labs(x="候選人", y="得票數", title="臺中市第3選舉區結果") +
  geom_text(aes(x=reorder(Candidate, Votes), y=Votes,
                label=sprintf("%s\n(%s)", scales::comma(Votes), scales::percent(Share))),
            size=5, hjust=-.1) +
  scale_y_continuous(labels=scales::comma, limits=c(0, max(tc3r$Votes) * 1.2)) +
  coord_flip()

Taichung 3rd District: Constituency Candidate Share

tc3_plr <- tc3_pl[, .(Votes=sum(Votes)), by=.(Party)]
tc3_plr <- tc3_plr[, Share:=Votes / sum(Votes)]
tc3_plr <- tc3_plr[Share > .01]
setorder(tc3_plr, -Votes)
tc3_plr <- tc3_plr[, Party:=factor(Party)]
ggplot(tc3_plr, aes(x=reorder(Party, Votes), y=Votes, fill=Party)) +
  geom_bar(stat="identity") +
  scale_fill_manual(values=c("darkorange", "darkgreen", "yellow", "dodgerblue", "green", "green3", "lightblue")) +
  labs(x="政黨", y="得票數", title="臺中市第3選舉區不分區政黨結果") +
  geom_text(aes(x=reorder(Party, Votes), y=Votes,
                label=sprintf("%s\n(%s)", scales::comma(Votes), scales::percent(Share))),
            size=4, hjust=-.1) +
  scale_y_continuous(labels=scales::comma, limits=c(0, max(tc3_plr$Votes) * 1.2)) +
  coord_flip() +
  theme(legend.position="none")

Taichung 3rd District: Party-List Share

ggplot(tc3_final, aes(x=Share, y=C_Share)) +
  geom_point(size=.5) +
  geom_smooth(method="lm") +
  labs(x="不分區政黨得票率", y="立委候選人得票率",
       title="臺中市第3選舉區不分區政黨得票率Ｘ候選人得票率") +
  facet_wrap(Party ~ Candidate, ncol=3, scales="free")

Taichung 3rd District: Party-List Share vs Constituency Candidate Share

Go Back to Top

References on R Packages

Berger, Susanne, Nathaniel Graham, and Achim Zeileis. 2017. “Various Versatile Variances: An Object-Oriented Implementation of Clustered Covariances in R.” Working Paper 2017-12. Working Papers in Economics; Statistics, Research Platform Empirical; Experimental Economics, Universität Innsbruck. http://EconPapers.RePEc.org/RePEc:inn:wpaper:2017-12.

Dowle, Matt, and Arun Srinivasan. 2019. Data.table: Extension of ‘Data.frame‘. https://CRAN.R-project.org/package=data.table.

Pebesma, Edzer. 2018. “Simple Features for R: Standardized Support for Spatial Vector Data.” The R Journal 10 (1): 439–46. https://doi.org/10.32614/RJ-2018-009.

Qiu, Yixuan, and authors/contributors of the included software. See file AUTHORS for details. 2019. Showtext: Using Fonts More Easily in R Graphs. https://CRAN.R-project.org/package=showtext.

Sievert, Carson. 2018. Plotly for R. https://plotly-r.com.

Slowikowski, Kamil. 2019. Ggrepel: Automatically Position Non-Overlapping Text Labels with ’Gplot2’. https://CRAN.R-project.org/package=ggrepel.

Wickham, Hadley. 2016. Ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York. https://ggplot2.tidyverse.org.

Zeileis, Achim, and Torsten Hothorn. 2002. “Diagnostic Checking in Regression Relationships.” R News 2 (3): 7–10. https://CRAN.R-project.org/doc/Rnews/.